CS7643 Deep Learning - Module 2 (Convolutional Neural Networks)¶
Lesson 5: Convolution and Pooling Layers¶
Convolution Layers¶
Limitation of Linear Layers¶
- Having fully connected layers all the way through isn't always the best choice.
- When input image has lots of layers, this could lead to overparameterization.
- Ex. 1024 (M) x 1024 (N) pixel image = M*N + bias = hundres of millions of parameter for just one layer.
- Overparameterization == overfitting and more data needed.
Locality of Features¶